The "Memory Foundation Model" released by the MemTensor team: adds a native memory module alongside the backbone attention, writing historical conversations into a fixed-size memory state, so raw context is no longer needed during inference.
Stars 432 (+9 / 24h)
Reasons
- Two versions: Nano is based on Qwen3.6-35B-A3B (total parameters 36B, active 3B), and Flash is based on DeepSeek-V4-Flash (total parameters 288B, active 17B); weights are hosted under the MemTensor organization on Hugging Face
- Officially claimed that at a history length of 128K, the persistent state saves approximately 66 times more storage compared to a full KV cache; on 11 memory benchmarks, the two models rank first and second among parametric memory methods
- The official documentation also states: for long histories requiring detailed evidence retrieval, text-based memory systems still maintain an overall lead
- Includes an example of an "experience-guided coding agent": retrieved experiences are stored in native memory and do not occupy task context; see the technical report on arXiv 2610.26825
Usage
- On a Linux server with NVIDIA CUDA, navigate to the inference directory and run uv pip install '.[server]'; install '.[client]' in the client environment
- Start the service with sh start_memt_nano.sh /models/MemT-1.0-Nano (the example uses two GPUs), then converse using MemTSession from memt_sglang; when enable_memory=True, conversations are written to memory
Risks
- The code is licensed under PolyForm Noncommercial 1.0.0, and the paper materials are under CC BY-NC-SA 4.0; neither permits commercial use
- Requires multi-GPU setups and full BF16 weights; the conversation examples in the README are illustrative, not actual run logs
- The number of new stars in the last 24 hours is very low, indicating limited attention